How to Tell if an Image Is AI Generated

To tell if an image is AI generated, you need to look for inconsistencies in details like lighting and shadows. You can also check the image metadata if it is available. AI pictures often look too perfect or have a plastic quality in the skin and lighting.
You should examine the finer details for discrepancies. Shadows might not align with the light sources. You might see distorted objects or areas that are blurred or overly sharp in ways that seem out of place. The scale of things and the shading often look unnatural.
AI models often struggle with coherent text. You might see misspelled words, gibberish, or inconsistent fonts in the picture. The background also gives clues. The further back something sits, the more the model stops committing to real geometry and turns it into a flat texture.
You can look for duplicate elements and a lack of realism. You might see identical background objects like trees or bushes that look exactly alike. You might also see repetitive facial features. Some AI systems like Google SynthID embed invisible watermarks that specific tools can detect.
Many models insert metadata into the image file, but platforms often strip this information when you upload the file. It is becoming harder to spot these images because the models are constantly improving. Common tells like distorted hands are becoming less frequent, and the artifacts are becoming more subtle every day.
Things to check
- Lighting and shadows Look for shadows that do not align with the light source or a plastic look.
- Text and fonts Check for misspelled words, gibberish, or inconsistent fonts.
- Background geometry Look at distant objects to see if they turn into flat texture instead of real shapes.
- Duplicate elements Check for identical trees, bushes, or people in the background.
- Metadata and watermarks Look for origin data or invisible watermarks like Google SynthID.

Inconsistencies and Anomalies
Metadata and Watermarking
The Evolving Challenge
Do you want to know about specific tools that can help detect AI-generated images?
Bottom line
To tell if an image is AI-generated, look for inconsistencies in details like lighting, shadows, text, and repeated elements, as well as metadata if available.
Community answers 27
What others in the community said:
Saw a comment in a random post where an user checked if the picture posted by op is ai generated or not , so I checked some of the pictures from here . All you have to is go to gemini and type /synthID with the image. It detects an invisible watermark of ai generated images
Sometimes there are some pretty obvious artifacts of artificial generation if you know what you are looking for. But unfortunately, those are getting a lot more subtle everyday, and before long it will be basically impossible to tell.
I usually can tell instantly when something is ai-slop, but i only know that beacuse of the feel the picture or the songs gives me; if someone where to ask me the clues catch ai-slop, i wouldn't know the answer.
Since piss filters and morphed hands are less common than before in ai-generation, what should i look for?
Some of the AI systems do stamp them with a label (like in the code, not the watermark), but not all. It really should be legislated.
Just working on improving image quality to get better details, and cleaner overall realism so the results look less ‘AI-generated’ and more like real photos for my app CosplayAI
odd reflections, wrong lighting, out of place scenery.
or, my good ole gut feeling
Found some interesting old photos from 1975 on twitter but when I posted them I was alerted that they might be AI generated.
I hate AI and would never help spread its shit over the internet, so I deleted the post even though I had no way of knowing whether they are indeed AI or not. I'd rather be on the safe side, you know.
Now, the question is; how do I check?
The photos I posted were so realistic, even with the appearance of wear and tear on some of them.
Any recommendations for sites or apps to do that?
Thanks and fuck AI.
I know their are some details that AI can't quite get it yet. But, I remember a tweet by some account with an image of a hot woman holding a gun, and somebody quote tweeted it with the same image, but with some distortion or something (it looked like static on old TVs) and claiming it was AI.
I wonder if it's actually a method to recognise AI generated images or is just Bullshit?
I follow a lot of video game subreddits and people will post cosplays and art all the time. Sometimes if it looks good, I’ll go to make a comment and the comment section will already be filled with “AI slop” “AI not your drawing” “this is AI not a cosplay”.
Then almost always someone will ask how do you know it’s AI and they’ll be like “look at the hand” “such and such has the wrong hair color” “it’s just clearly fake”.
I’m not a fan of AI so I’d like to be able to figure out what is fake or not too. Or are people just calling everything with a minor mistake (such as a drawing where the character has a different hair color) AI now?
I often can tell by how the lighting is done. It makes the picture seem rather... plastic?
AI makes decisions that a human artist would never make. If it looks like someone spent days on photo realistic rendering, but they didn't take a few seconds to make sure that what they're drawing makes physical sense, you just know that no human would do that. No human would waste that much time when there's an obviously weird looking error that you can spot right away.
There's also a point where no human artist would spend as many hours as it would take to do an incredibly detailed and rendered drawing, when the subject of the drawing is just a stupid thing that no one would waste that much time on.
Gemini and some others specifically include markers for it, so you can use their tools to at least check for their own stuff.
While judging an image is there something that I can watch out for that signals that it’s AI generated? Does AI generated image carry a ‘signature’ to identify it is AI?
Shadows facing toward the light source, and objects casting no shadows in bright light.
Two or more background scenery pieces (i.e., trees, bushes, people, et cetera) looking exactly alike.
Multiple faces looking exactly alike (i.e., features, expressions, et cetera).
Signs with random letters that spell out nothing in any known language.
No. There is no filter you can run an image through to magically detect AI. Like all methods of spotting AI what works today likely wont work in the future or work broadly in the past.
Google has created SynthID which are watermarks placed into generativly created AI content via their models.
This serves to help ID generative AI but also allows companies to know if something came out of their specific model.
The images usually look to perfect, to rounded and smooth. Also the scale of thing, shadows and shading are not natural. An in many cases u will find artifacts, like weird hands.
the fact that this thread can't agree is kind of the answer. couple years ago it was hands and text that gave it away, now those are mostly fixed. the one that still gets me is deep backgrounds. the further back something sits the more the model stops committing to real geometry and it just turns into texture. that's the first place i look now.
All of them except first one.
Many of the models insert metadata, but when the image is uploaded to various places the metadata is often stripped. Detection is multi pronged and meta is one of the first steps in detection. If no meta is found you move to next methods.
It’s in the minute details. There’s an image floating around of two people in Minnesota sitting on a Nissan Sentra with automatic guns. Perfect example of a ai influenced image. If you can find it look at the car model badge on the back trunk. It’s messed up. Also the word backrack in the back part of the image is wrong and inconsistent. One c is different than the other c in the word. It’s the subtleties that will show it’s fabricated with ai. 🤖
Googles trying SynthID to add water marks to text and images but how it actually works sounds complicated. Metadata can be stripped, visible water marking is easily removed. Right now the generation is still a bit flawed so there are some mistakes. But as it improves these we be harder and harder to catch. Likely it will become a mater of trust, who put out the image and do I trust them?
- Digital watermarks: Some AI systems, like Google's SynthID, embed invisible watermarks that can be detected by specific tools.
- Rapid improvement: AI models are constantly improving, making it increasingly difficult to spot AI-generated images as common
Replies (0)
No replies yet. Be the first to reply.